集成PCA和LSTM神经网络的浸润线预测方法

为预防尾矿库溃坝事故,挖掘尾矿库在线监测系统的有效信息,提高浸润线预测精度,构建基于主成分分析(PCA)和长短期记忆(LSTM)神经网络的浸润线预测模型;以陈坑尾矿库为例,引入皮尔森(Pearson)相关系数和变量组合法,确定模型输入为预测前3天的待测点浸润线位置、相邻周边2点浸润线位置、库水位、坝体纵向位移和降雨量等18个特征量;利用PCA消除输入变量间的数据冗余,并采用LSTM神经网络预测未来3天的浸润线位置。结果表明:基于PCA和LSTM神经网络的浸润线预测方法具有较高的预测精度,平均绝对误差为0. 011,决策系数为0. 805,且能实现不同降雨工况下尾矿库浸润线的稳定预测。

A PCA-LSTM neural network-integrated method for phreatic line prediction

In order to prevent dam-breaking accidents of tailings ponds,to excavate effective information of online monitoring system and improve prediction accuracy of phreatic lines,a prediction model was set up based on PCA and LSTM neural network. Then,with Chenkeng tailings pond as an example,Pearson correlation coefficient and variable combination method were introduced to determine 18 features of model inputs,including location of phreatic line of measuring point in the first three days,location of two adjacent surrounding saturation lines,water level of ponds,longitudinal displacement of dam body and rainfall. Finally,PCA was used to eliminate data redundancy between input variables,and LSTM neural network was applied to predict location of phreatic line for the next three days. The results show that PCALSTM neural network-based method presents higher predication accuracy with an average absolute error of0. 011 and a decision coefficient of 0. 805. And it can achieve stable prediction of phreatic lines for tailings ponds under different rainfall conditions.